Professional Documents
Culture Documents
Disclaimer: This paper should not be reported as representing the views of the European Central Bank
(ECB). The views expressed are those of the authors and do not necessarily reflect those of the ECB.
Abstract: This paper compares within-sample and out-of-sample fit of a DSGE model with ra-
tional expectations to a model with adaptive learning. The Galí, Smets and Wouters model is the
chosen laboratory using quarterly real-time euro area data vintages, covering 2001Q1–2019Q4. The
adaptive learning model obtains better within-sample fit for all vintages used for estimation in the
forecast exercise and for the full sample. However, the rational expectations model typically predicts
real GDP growth better as well as jointly with inflation. For the marginal inflation forecasts, the
same holds for the inner quarters of the forecast horizon, while the adaptive learning model predicts
better for the outer quarters.
Keywords: Bayesian inference, CRPS, euro area, forecast comparison/evaluation, log score, real-
time data.
JEL Classification Numbers: C11, C32, C52, C53, E37.
Estimated dynamic stochastic general equilibrium (DSGE) models are regularly used by many central
banks and other policy-oriented organizations to analyse macroeconomic conditions and to assess
future developments. Given the micro-foundations and optimization-based behavior of economic
agents in such models, an important reason for their popularity is that they facilitate structural in-
terpretations of the macroeconomic environment. Expectations are typically assumed to be rational
such that the joint probability distribution of the model variables is fully consistent with the model
and policy experiments are therefore not subject to the Lucas critique. Furthermore, the rational
expectations (RE) assumption provides strong cross-equation restrictions on the resulting stochastic
processes which describe the model variables and which help to better identify many of the parame-
ters. Bayesian inference is frequently used for estimation, while calibration of some parameters that
may be hard to identify from the macro data concerns relatively few parameters in the benchmark
models, such as the well-known Smets and Wouters model, and where instead micro data may be
consulted.
The main contribution of the current paper is the real-time density forecast comparison and
evaluation of a benchmark DSGE model under RE and adaptive learning (AL). To my knowledge,
such an empirical study is not available in the literature and, as a complement to the out-of-sample
investigation, the models are also compared recursively within-sample to learn if the fit of a DSGE
model subject to AL is better (or worse) than under RE for the euro area data. The selected DSGE
model is the euro area version of the Galí, Smets and Wouters model, which extends the Smets and
Wouters model to incorporate unemployment. The real-time database for the euro area is considered
for the forecast exercise, covering backcasts, nowcasts and up to eight-quarter-ahead forecasts over
the forecast sample 2001Q1–2019Q4. The density forecasts are compared using the log score as well
as the continuous ranked probability score (CRPS) and its multivariate version, the energy score,
while the predictive distributions are evaluated using formal tests based on the standard probability
integral transform and, in the case of multivariate distributions, Box’s density ordinate transform.
Concerning within-sample fit, the AL model has a considerably greater log marginal likelihood
than the RE model for the full estimation sample, i.e., when both models are estimated until 2019Q4.
This confirms the findings from US data that DSGE models subject to AL fit the estimation data
better than under RE. Furthermore, the two models are also compared within-sample for each Q1
vintage over the real-time forecast sample, 2001–2019. Again, the AL model obtains bigger log
marginal likelihood values for each such vintage.
Turning to the forecast comparison exercise, the RE model overall predicts real GDP growth
more accurately than the AL model from a point forecast perspective. Both models over-predict
real GDP growth on average and the AL model substantially for the four-quarter-ahead to eight-
quarter-ahead forecasts. The inflation forecast paths reflect a sharp difference between the dynamic
behavior of the two expectation mechanisms. The paths from the RE model are typically upwarding
sloping with a concave curvature during the comparison sample and result in an over-prediction of
inflation on average, while the AL model paths tend to be u-shaped and lead to an under-prediction
Estimated dynamic stochastic general equilibrium (DSGE) models are regularly used by many central
banks and other policy-oriented organizations to analyse macroeconomic conditions and to assess
future developments; see, e.g., Lindé, Smets, and Wouters (2016) and Christiano, Eichenbaum, and
Trabandt (2018) for recent surveys. Given the micro-foundations and optimization-based behavior
of economic agents in such models, an important reason for their popularity is that they facilitate
structural interpretations of the macroeconomic environment. Expectations are typically assumed to
be rational such that the joint probability distribution of the model variables is fully consistent with
the model and policy experiments are therefore not subject to the Lucas critique. Furthermore, the
rational expectations (RE) assumption provides strong cross-equation restrictions on the resulting
stochastic processes which describe the model variables and which help to better identify many
of the parameters. Bayesian inference is frequently used for estimation, while calibration of some
parameters that may be hard to identify from the macro data concerns relatively few parameters in
the benchmark models, such as the well-known Smets and Wouters (2007) model, and where instead
micro data may be consulted.
The main contribution of the current paper is the real-time density forecast comparison and
evaluation of a benchmark DSGE model under RE and adaptive learning (AL) using the Slobodyan
and Wouters (2012a) approach. To my knowledge, such an empirical study is not available in the
literature and, as a complement to the out-of-sample investigation, the models are also compared
recursively within-sample to learn if the fit of a DSGE model subject to AL is better (or worse)
than under RE for the euro area data. The selected DSGE model is the euro area version of the
Galí, Smets, and Wouters (2012, SWU) model, which extends the Smets and Wouters (SW) model
to incorporate unemployment; see Smets, Warne, and Wouters (2014) and McAdam and Warne
(2019) for details on the euro area version of the SWU model. The real-time database for the euro
area, introduced by Giannone, Henry, Lalik, and Modugno (2012), is considered for the forecast
exercise, covering backcasts, nowcasts and up to eight-quarter-ahead forecasts over the forecast
sample 2001Q1–2019Q4. The density forecasts are compared using the log score as well as the
continuous ranked probability score (CRPS) and its multivariate version, the energy score, while the
predictive distributions are evaluated using formal tests based on the standard probability integral
transform and, in the case of multivariate distributions, Box’s (1980) density ordinate transform.
The forecasting methodology for DSGE models is discussed in Del Negro and Schorfheide (2013),
which also provides illustrations on the empirical performance of such models relative to survey data,
professional forecasts and reduced form models for the US; see Adolfson, Lindé, and Villani (2007)
and Christoffel, Coenen, and Warne (2011) for discussions on methodology and empirical evidence
from the euro area. The point and density forecasting performance of DSGE models compares well
with reduced form benchmark models, such as BVARs, especially over the medium term. This
general finding is also supported by Warne, Coenen, and Christoffel (2017) for a larger dimension
DSGE models are usually estimated and analysed under the assumption that expectations are formed
rationally and therefore model consistent. The structural form of a log-linearized DSGE model is
represented by:
H−1 zt−1 + H0 zt + H1 Et zt+1 = Dηt , t = 1, 2, . . . , (1)
4
Forecast comparisons for RE and AL models are otherwise very rare in the literature. In a recent paper, Carvalo,
Eusepi, Moench, and Preston (2022) develops a Rotemberg-pricing model with imperfect information to describe
long-term trends in inflation expectations and shows that it can explain developments in long-term forecast data from
US professional forecasters better than under RE. Elias (2022) estimates a small-scale new Keynesian model with
heterogeneous Euler equation AL on US data and finds that a model with heterogeneous AL fits the data better than
a model with homogeneous AL. Point forecasts of the output gap and inflation made by the three types of agents in
the model are compared over the pre-Great Moderation, Great Moderation and pre-Great Recession periods and the
RMSEs suggests heterogeneity matters. Still, these exercises are not carried out with an RE version of the model.
where F and B satisfy the cross equation restrictions B = (H0 +H1 F )−1 D and H−1 +H0 F +H1 F 2 =
0. When estimating such a DSGE model, the solution corresponds to the state equation, while the
measurement equation is given by
yt = µ + H ′ zt + wt , (3)
5 See, for instance, Eusepi and Preston (2018b, Section 4.2) and references therein regarding the impact of the so-called
anticipated utility approach relative to the Euler-equation approach for the DSGE model solution(s) under learning
or imperfect information.
zt = µt + Ft zt−1 + Bt ηt , (4)
The well-known Smets and Wouters (2007) (SW) model is founded on a continuum of utility-
maximizing households and profit-maximizing intermediate-good-producing firms who, respectively,
supply labor and intermediate goods in monopolistic competition and set wages and prices. Final-
good-producers use these intermediate goods and operate under perfect competition. The model
incorporates several real and nominal rigidities, such as habit formation, investment adjustment
costs, variable capital utilization and Calvo staggering in prices and wages. The monetary author-
ity follows a Taylor-type rule when setting the nominal interest rate. There are seven stochastic
processes: a TFP shock; a price and a wage markup shock; a risk premium (preference) shock; an
exogenous spending shock; an investment-specific technology shock; and a monetary policy shock.
The observed variables of the euro area version in McAdam and Warne (2019) are: real GDP, real
private consumption, real total investment, total employment, real wages, the GDP deflator and the
short-term nominal interest rate.
The euro area version of the Galí et al. (2012) extension of the SW model is presented in Smets
et al. (2014). It explicitly provides a mechanism for explaining unemployment. This is accomplished
by modelling the labor supply decisions on the extensive margin (whether to work or not) rather
than on the intensive margin (how many hours to work). As a consequence, the unemployment
rate is added as an observable variable, while labor supply shocks are allowed for. The extension
is motivated by the critique that the SW model cannot properly identify wage markup shocks; see
Chari, Kehoe, and McGrattan (2009). The inclusion of observable unemployment into the wage
equation can help to overcome that problem. Henceforth and following McAdam and Warne (2019),
this extension is called the SWU model below.
Smets et al. (2014) conducts a point forecast study with this model and a few benchmarks utilizing
the real-time database (RTD) of the euro area for the sample 2001Q1–2011Q4; see Giannone et al.
(2012).7 McAdam and Warne (2019) also make use of the SWU model to compare its real-time
density forecasting properties with the SW model as well as with an extension of the latter subject to
7 A somewhat more detailed presentation of the log-linearized SWU model is available in the Online Appendix to
McAdam and Warne (2019), while the Appendix to McAdam and Warne (2018) contains even further details, including
the flexible price part of the model.
The construction of the RTD combined with the AWM database is discussed in some detail by
Smets et al. (2014) and the extension until 2015Q4 is covered in McAdam and Warne (2019). For
the current study, the data vintages have been extended until 2020Q4 and the variables involved are
real GDP, real private consumption, real total investment, total employment, the GDP deflator, real
wages, the nominal 3-month Euribor interest rate and the unemployment rate. These variables are
transformed exactly as in the above two studies. For some further details on the data, see Table B.1
in the Appendix.
One important difference for the additional vintages concerns the total investment data. Due to
accounting issues regarding mainly multinational corporations in Ireland, but also for the Nether-
lands, some periods display extreme spikes. This is shown in Figure 1 for real total investment
growth for the full sample discussed in Section 4.1. It is noteworthy that the numbers in the time
periods 2015Q2–Q3, 2017Q2–Q3, and 2019Q2–Q4 are very big in an absolute sense compared with
the overall volatility of the variable. Moreover, the growth rate is unusually high (low) in Q2 (Q3),
consistent with accounting related data problems. Finally, these outliers first appears in the 2018Q3
vintage and, hence, the number of affected vintages is fairly small.
Although the forecasting exercise does not attempt to predict investment growth, such outliers
can influence the estimation parameters in a recursive exercise. In this paper, the outlier values have
been replaced with the sample mean, but also the alternatives of keeping the data for those periods
as well as treating the outlier periods as missing data have been examined. The choice between
The forecasting exercise in Section 5 uses annual revisions for the actual (true) values of the variables
with the consequence that the last vintage used for prediction is 2019Q4. In this subsection the
parameter estimates of the SWU model under RE and AL are discussed. The focus is on the full
sample 1979Q4–2019Q4 using the latest update (2018) of the Area-Wide Model (AWM) database
combined with the RTD vintage 2020Q3.9 The third quarter corresponds to the time period when
the AWM database has typically been frozen with data until the end of the previous year and, hence,
mimics an update from 2020.
The estimated SWU model in Smets et al. (2014) is based on the assumption of log utility
in consumption and this variant is also used by McAdam and Warne (2019). This means that the
inverse elasticity of intertemporal substitution for constant labor, denoted by σc , is equal to 1. When
estimating the SWU model under AL this parameter is instead estimated and the prior distribution
is given by σc ∼ N (1, 0.252 ) so that it is centered around unity. Furthermore, and as in McAdam
and Warne (2019), the MA parameters of the price and wage markup shocks are set to 0 under RE
as well as under AL. Finally, and as mentioned above, the baseline RE model uses the flexible-price
output gap and is therefore the same specification as in McAdam and Warne (2019), while under
AL the output gap is defined as in Slobodyan and Wouters (2012a).10
Concerning the AL baseline model, the scale parameters are calibrated to σr = 0.03 and σε =
0.003, as in Slobodyan and Wouters (2012a), while the prior distribution of the autoregressive
parameter ρ is a standard beta with mean 0.25 and standard deviation 0.1 in the baseline version.11
This informative prior implies that the belief coefficients return fairly quickly to their steady-state
values. Moreover, the forward looking variables are eight and given by consumption, investment,
inflation, hours worked, the value of the capital stock, the rental rate of capital, real wages and
employment. The PLM for each of these variables is equal to a constant and two own lags with
time-varying parameters. Finally, the autoregressive parameter of the price markup shock process
is calibrated to zero, such that this process is i.i.d. in the baseline model.
A selection of parameter estimates of the baseline models is shown in Table 1, which also provides
details on the estimation procedure.12 The focus is on the wage and price markup shock related
8 Estimating the AL baseline model with the outliers has an impact on investment related parameters. Both the
elasticity of the capital adjustment and of the capital utilization cost functions drop as does the estimated capital share.
Furthermore, the estimated persistence parameter of the investment-specific technology shock falls, while the estimated
shock standard deviation increases. For the RE baseline version the use of the outliers have a somewhat smaller effect
on the two elasticities, while the impacts on the capital share and investment-technology shock parameters are very
present. In addition, the steady-state growth rate drops for the RE baseline model.
9 The observed variables are plotted in the Appendix, Figure C.1, where total investment growth is plotted with the
sample mean correction for the selected outlier periods.
10 The prior distributions of all parameters are shown in the Appendix, Tables B.3–B.4.
11 For comparisons, keep in mind that a standard uniform distribution is equivalent to a standard beta distribution
√
with mean 1/2 and standard deviation 1/ 12 ≈ 0.28868.
12
All parameter estimates are available in the Appendix, Tables B.5–B.6.
13 The log marginal likelihood for the case when ρ = 0 (the belief coefficients are white noise around the steady-state)
is estimated at -423.79. The posterior parameter estimates are generally similar to the models where ρ is estimated
Based on the evidence in the previous section, there is empirical support for the view that learning
improves the within-sample fit of DSGE models. This is in line with empirical studies on US data,
such as Milani (2007) and Slobodyan and Wouters (2012a,b). In this subsection, the within-sample
fit of the baseline SWU models under RE and AL are estimated on vintages from the RTD that are
used in the forecasting exercises in Section 5. Following McAdam and Warne (2019), the models are
re-estimated on an annual frequency using the Q1 vintage with a sample covering the vintages from
2001 to 2019. The selected frequency mimics, under the best of circumstances, how often structural
models are re-estimated in practise at policy institutions.
The recursive estimates are displayed in Figure 2. In the upper panel posterior-mode-based
goodness-of-fit estimates are shown, while in the lower panel two estimators of the log marginal
likelihood, the modified harmonic mean and the Laplace approximation, are plotted. Turning first
to the upper panel, the estimated log likelihood (solid lines), the log posterior kernel (dash-dotted
lines) and the Laplace approximation of the log marginal likelihood (dashed lines) are plotted with
red lines for the RE model and blue for the AL model.14 It is noteworthy that the difference between
the blue and the red lines for each statistica over the sample is visually very similar. This suggests
that mainly the data are driving these results, while the influence of the prior is negligible. Moreover,
the AL model always has a greater value and the gap to the RE model is slowly increasing over time.
In the lower panel we find both the Laplace approximation of the log marginal likelihood from the
upper panel with dashed lines, and the modified harmonic mean (MHM) estimator using a selection
of the post-burnin sample of draws from the posterior distributions. It is striking how small the
differences between the two estimators are, with mainly visible differences for the AL model. The
finding from the full sample that the AL model fits the euro area data better than the RE model is
therefore also supported by the recursive estimates from the real-time sample 2001–2019.
In this Section four topics are covered. First, we examine the point forecasts of the RE and AL
models using the recursive posterior mean paths. Second, we turn to the density forecast where two
scoring rules are considered: (i) the standard log score, and (ii) the continuous rank probability score
for univariate forecasts and the energy score for multivariate forecasts; see, for instance, Gneiting
and Raftery (2007) for a survey on scoring rules and Gneiting and Katzfuss (2014) for a review
on probabilistic forecasting. While both these topics cover forecast comparisions, the third topic is
concerned with forecast evaluation using the probability integral transform (PIT) in the univariate
and Box’s density ordinate transform (BOT) in the multivariate case. Finally, we consider some
sensitivity analysis.
with the exception of the wage markup shock persistence parameter, which is substantially lower for the ρ = 0 case
with a posterior mean of 0.22.
14
The inverse Hessian used for computing the Laplace approximation is based on a finite difference estimator available
in the well-known software Dynare.
The point forecast is given by the mean of the predictive distribution and is estimated by averaging
the point forecast conditional on the parameters over the posterior draws; see, e.g., McAdam and
Warne (2019, Section 5.2). These calculations have been performed using 10,000 equally spaced
posterior draws from the 500,000 post burn-in draws of the random-walk Metropolis sampler, where
the size of the burn-in sample is 250,000 draws. The resulting “spaghetti” plots are shown in Figure 3
for real GDP growth in the upper panel and GDP deflator inflation in the lower panel. The forecasts
from the baseline RE model are shown to the left with red solid lines and the forecasts from the
baseline AL model to the right with blue solid lines.16 The actual values are given by the black
solid line, while the black dashed line provides the recursive posterior mean estimates of steady-
state real GDP growth and GDP deflator inflation, respectively. Similarly, the mean forecast errors
(actual values minus prediction) and the Diebold-Mariano tests for equal mean square forecast errors
(MSFE) are provided in Table 2.17
Turning first to the real GDP growth forecasts in the upper panel of Figure 3, the forecasts from
the AL model are overall flatter and are typically greater than the ones from the RE model. This
is confirmed in Table 2 where the mean errors are negative and bigger in absolute terms for the AL
model. In addition, the difference in the mean errors is overall increasing with the horizon. This is
further supported by the Diebold-Mariano tests, where the small cdf-values for the longer horizons
indicate that the RE model forecasts real GDP growth better than the AL model.
The point forecasts of inflation in the lower panel exposes the path contrasts between the RE
and AL models. While the former are typically strongly upward-sloping, concave and tend to over-
predict inflation, the latter paths are often u-shaped and fairly close to the actual values for the
outer quarters of the horizon. This visual impression is confirmed by the results in Table 2 where
the RE model has negative mean errors which increase in absolute terms with the forecast horizon.
By contrast, the AL model under-predicts inflation but the mean errors are approaching zero as
15
The ragged edge of the real-time data is displayed in Table B.2 of the Appendix.
16
It should be noted that the forecasts for the AL model are computed under the assumption that the belief coefficients
are held fixed over the forecast horizon. This is common in the learning literature; see, e.g., Eusepi and Preston (2018b)
for discussions on this and related matters. The dynamic responses to an impulse/shock in any given time period are
also computed under the assumption of fixed beliefs while updating of the beliefs occurs once the time period for the
impulse moves forward one period.
17
The prediction errors for the individual horizons of real GDP growth and inflation are plotted in the Appendix,
Figures C.2–C.3.
Density forecasts are compared across models using a scoring rule and below two such rules are
utilized, serving different aspects of the predictive distribution. A scoring rule is said to be proper if
a forecaster who maximizes the expected score provides the true subjective predictive distribution,
and it is said to be local if the rule only depends on the predictive density at the realized value of
the predicted variables. In this section the log score is used and, being the only proper local scoring
rule, is equal to the sum (over the vintages in the comparison sample) of the log predictive likelihood
evaluated at the actual value. It is estimated below as described by, for example, Warne et al. (2017)
and McAdam and Warne (2019). The next section will focus on a proper scoring rule which observes
the cumulative predictive density and therefore departs from the restriction of a local scoring rule.
18
See Diebold (2015) for discussions on the use and abuse of the Diebold-Mariano test and more generally on the
usefulness of out-of-sample forecast comparisons.
19
The overall dynamic pattern of the impulse responses under RE does not change much from the choice of estimates.
For instance, the real GDP impulse responses in the baseline RE model at its posterior mode estimates are approxi-
mately the same, while the inflation responses have the same shape as for RE in Figure 4, but the initial response is
much weaker at around −0.015.
20
McAdam and Warne (2019, Table 8) finds that for the euro area sample until 2014Q4 the log predictive likelihood
under RE for the SWU model is well approximated by a Gaussian distribution with mean and variance taken from
the predictive distribution.
21
The estimated log predictive likelihood values of the RE (AL) model in 2016 are for each quarter respectively −0.726
(0.012), −1.108 (−0.104), −0.942 (−0.131) and −0.243 (−0.140). The numerical standard errors of these estimates
are very small and around 0.01 for the RE model and less than half of that for the AL model.
22
Presumably the ELB lies close to the path for the short-term interest rate when it is close to zero or negative.
Such a hard condition of the actual values avoids obtaining paths for the interest rate that lie below the actual values.
However, it also disallows interest rate paths with values above the actual values and where such interest rate paths
are expected to be linked with lower projected inflation paths. Hence, this experiment will most likely over-estimate
inflation relative to imposing the ELB, but has the advantage of being straightforward to compute and avoids the
extremely low interest rate paths.
The log score uses only information about the predictive density at the actual value while, for
instance, values near it and that have a high likelihood are ignored. Scoring rules based on the
predictive cdf allows for a wide range of possible values and therefore provide a more comprehensive
measure of density forecast performance than the log score. Furthermore, it is shown by Krüger,
Lerch, Thorarinsdottir, and Gneiting (2021) that the theoretical conditions for the log score to be
consistent when using MCMC draws of the parameters are stronger than for some alternative scoring
rules. For univariate forecasts, they show that the continuous ranked probability score (CRPS) is
consistent under weaker conditions. In addition, the CRPS is proper and, under certain conditions,
even strictly proper; see Gneiting and Raftery (2007) for details. This scoring rule is often expressed
as
1 h
(c)
i h
(o)
i
CRPS(h, i) = EP D yi,T +h − yi,T +h − EP D yi,T +h − yi,T +h , i = 1, . . . , np , (5)
2
(c)
where np is the number of predicted variables, yi,T +h and yi,T +h are independent copies of the
predicted variable i in time period T + h given the information at T with (posterior) predictive
(o)
distribution P D, and with actual values yi,T +h for the h-quarter-ahead forecasts. For a given vintage
T , this scoring rule can be consistently estimated using P simulated forecasts paths based on the
posterior draws of the parameters by
P P P
1 X X (j1 ) (j2 )
1 X
(j) (o)
CRPS(h, i) = y − y − y − y i,T +h , i = 1, . . . , np , (6)
i,T +h i,T +h i,T +h
2P 2 P
j1 =1 j2 =1 j=1
(j)
where yi,T +h is the value of variable i in simulated path j at T + h. For the cumulative density
forecast comparisons, the CRPS is given by the sum of (6) over the 76 vintages at hand.
A multivariate extension of the CRPS, called the energy score, was introduced by Gneiting and
Raftery (2007) and applied by, e.g., Gneiting, Stanberry, Grimit, Held, and Johnson (2008). It can
be expressed as
1 h
(c)
i h
(o)
i
ES(h, s) = EP D
y1:np ,T +h − y1:np ,T +h
− EP D
y1:np ,T +h − y1:np ,T +h
, (7)
2
√
where kxk = x′ x is the Euclidean norm, and y1:np ,T +h is a vector with the np predicted variables.
This expression can be consistently estimated using P paths from the predictive distribution as
P P P
1 X X
(j1 ) (j2 )
1 X
(j) (o)
ES(h, s) = − y − − y 1:np ,T +h
. (8)
2P 2
y 1:np ,T +h 1:np ,T +h
P
y 1:np ,T +h
j1 =1 j2 =1 j=1
It can be noted that the CRPS for a deterministic forecast system reduces to minus the absolute
point forecast error. Hence, the CRPS (and ES) encompasses both probabilistic and deterministic
scoring rules.
The probability integral transform (PIT) has long been used to assess if a forecasting model is well
calibrated. An early paper which considered this idea for density forecasting purposes in econometrics
is Diebold, Gunther, and Tay (1998), but it has earlier been emphasized by Dawid (1984). Rosenblatt
23 The recursive average CRPS values for real GDP growth and inflation, respectively, are shown in Figures C.6–C.7
of the Appendix.
is independent and uniformly distributed on the unit interval, where Fi (·) is the cdf and YT is the
information set available at T . Smith (1985) further noted that zi,T +1|T = Φ−1 (πi,T +1|T ), where
Φ(·) is the cdf of the normal distribution, is i.i.d. N (0, 1).
Amisano and Geweke (2017) construct a test statistic based on the normality property of the
inverse cdf; details are available in their Online Appendix. Specifically, let
(j) (o) (j) (j)
πi,T +1|T = Φ yi,T +1 µi,T +1|T , σii,T +1|T , j = 1, . . . , N
(j)
where µi,T +1|T is the one-quarter-ahead point forecast of predicted variable i = 1, . . . , np using the
(j)
j:th posterior draw of θ, the vector of estimated parameters, while σii,T +1|T is the one-quarter-ahead
forecast standard deviation of variable i using θ (j).
Next, the Monte Carlo average of the N values of the uniform variable is taken such that
N
1 X (j)
πi,T +1|T = πi,T +1|T , (9)
N
j=1
while
zi,T +1|T = Φ−1 πi,T +1|T .
(10)
Under the assumption that the model forecasts are well calibrated, the variable zi,T +1|T is normally
distributed with zero mean and unit variance. This assumption is tested in Amisano and Geweke
(2017) using the first q moments and p lags of the zi,T +1|T process with a test statistic which is
asymptotically χ2q+p .
The uniformity of the PIT does not hold for multivariate forecasts; see Genest and Rivest (2001).
As pointed out by Gneiting et al. (2008), an option is then to utilize the Box density ordinate
transform (BOT). The BOT was, e.g., proposed by Box (1980), and is defined as
h i
(o)
πT +1|T = 1 − Pr p y1:np ,T +1 YT ≤ p y1:np ,T +1 YT .
If y1:np ,T +1 is distributed as p(·) and this density is continuous, then πT +1|T is standard uniform.
For example, if the density p(·) is Gaussian with mean vector µ and covariance matrix Σ, then
′
(o) (o)
πT +1|T = 1 − χ2np y1:np ,T +1 − µ Σ−1 y1:np ,T +1 − µ . (11)
Using the Kolmogorov-Smirnov test with the SWU model under RE for the euro area RTD
McAdam and Warne (2019, Table 8) finds that the empirical distribution of the predictive likelihood
is not statistically different from a normal distribution with predictive mean and covariance matrix
estimated with Monte Carlo integration and evaluated at the actual value. These tests have been
recomputed with the extended real-time sample under the RE and the AL baseline models and the
conclusion is confirmed also for these cases; see Table B.7 in the Appendix. In the current study, the
BOT expression in (11) is therefore used to estimate πT +1|T for the multivariate case. Furthermore,
The PLM for the adaptive learning case has so far been assumed to be an AR(2) process with a
constant term for each forward looking variable. This simple PLM was also used by Slobodyan and
Wouters (2012a), but given the relatively poor forecasting performance of the AL model compared
with the RE version an important factor for this may be the assumed PLM. Theoretically there is
little guidance on the selection of a PLM, but from a time series perspective we may consider VARs.
Such a PLM is, however, very cumbersome from a computational perspective and is, as noted by
Slobodyan and Wouters (2012a), likely to suffer from overfitting. It may therefore be useful to
investigate the economic model for simpler suggestions.
With this in mind, the alternative PLM considered in this section uses the AR(2) process with
a constant term as a basis and adds at most one variable to each one of the eight processes and
limited to the first lag. Specifically, for real private consumption the lag of the nominal interest
rate is included, for real investment the lag of the real value of the capital stock is used, inflation
is linked to the lag of unemployment, hours worked makes use of the lag of employment, the rental
rate of capital considers the lag of capital services used in production as potentially informative, real
wages to the lag of inflation and, finally, employment to the lag of real wages. The only process for
a forward looking variable which remains unchanged is the real value of the capital stock.25
For the results on the average log scores of the joint real GDP growth and inflation case we
return to Figure 5. Generally, the alternative PLM appears to provide some improvement over the
benchmark PLM, especially for the nowcasts and the one-quarter-ahead forecasts. While the average
log scores of the two- and three-quarter-ahead forecasts are similar, the benchmark PLM tends to
forecast better until the Great Recession for the four-quarter-ahead to eight-quarter-ahead forecasts,
while the alternative PLM shows some improvement over the benchmark PLM thereafter and up to
24 Histograms of the π estimates for the PITs are provided in the Appendix, Figures C.8–C.9.
25
Note that this alternative PLM is similar to one of the alternatives discussed in Slobodyan and Wouters (2012a,
Section V.A).
This paper compares and evaluates real-time density forecasts of real GDP growth and inflation for
the euro area using an unemployment-based extension of the well-known Smets and Wouters (2007)
model, which supports either rational expectations (RE) or adaptive learning (AL), as modelled in
Slobodyan and Wouters (2012a). The forecast comparison sample begins with the first real-time
database vintage in 2001Q1 and ends in 2019Q4, thereby covering the period of wage moderation,
the global financial crisis, the Great Recession that followed, the European sovereign debt crisis,
and the period in which policy rates are close to the effective lower bound (ELB). The selection of
actual values of the observed variables follows “good practise” in the real-time literature and therefore
takes the annual revisions such that vintages until 2020Q4 are also required. This means that noisy
early estimates are avoided, such as first releases, but also that major changes to the computational
methodology are less likely to influence the outcome of the forecast analysis, as may be expected for
the euro area if the latest vintage would instead be used for the actual values; see, e.g., Croushore
and Stark (2001) and Croushore (2006, 2011) for discussions on using real-time data.
Concerning within-sample fit, the AL model has a greater log marginal likelihood than the RE
model by about 22 log-units for the full estimation sample, i.e., when both models are estimated
until 2019Q4.26 This confirms the findings from US data by, e.g., Milani (2007) and Slobodyan and
Wouters (2012a) that DSGE models subject to AL fit the estimation data better than under RE.
Furthermore, the two models are also compared within-sample for each Q1 vintage over the real-time
forecast sample, 2001–2019. Again, the AL model obtains larger log marginal likelihood values for
each such vintage.
Turning to the forecast comparison exercise, the RE model overall predicts real GDP growth
more accurately than the AL model from a point forecast perspective. Both models over-predict
real GDP growth on average and the AL model substantially for the four-quarter-ahead to eight-
quarter-ahead forecasts. The inflation forecast paths reflect a sharp difference between the dynamic
behavior of the two expectation mechanisms. The paths from the RE model are typically upwarding
sloping with a concave curvature during the comparison sample and result in an over-prediction of
inflation on average, while the AL model paths tend to be u-shaped and lead to an under-prediction
26
In terms of the base-10 log-scale, 22.07 natural log-units corresponds to about 9.58 log-10 units.
Adolfson, M., Lindé, J., and Villani, M. (2007), “Forecasting Performance of an Open Economy
DSGE Model,” Econometric Reviews, 26(2), 289–328.
Amisano, G. and Geweke, J. (2017), “Prediction Using Several Macroeconomic Models,” The Review
of Economics and Statistics, 99(5), 912–925.
Amisano, G. and Giacomini, R. (2007), “Comparing Density Forecasts via Weighted Likelihood
Ratio Tests,” Journal of Business & Economic Statistics, 25(2), 177–190.
Bańbura, M., Giannone, D., and Reichlin, L. (2010), “Large Bayesian Vector Auto Regressions,”
Journal of Applied Econometrics, 25(1), 71–92.
Bernanke, B. S., Gertler, M., and Gilchrist, S. (1999), “The Financial Accelerator in a Quantitative
Business Cycle Framework,” in J. B. Taylor and M. Woodford (Editors), Handbook of Macroeco-
nomics, volume 1C, 1341–1393, North Holland, Amsterdam.
Box, G. E. P. (1980), “Sampling and Bayes’ Inference in Scientific Modelling and Robustness,”
Journal of the Royal Statistical Society Series A, 143(4), 383–430.
Carvalo, C., Eusepi, S., Moench, E., and Preston, B. (2022), “Anchored Inflation Expectations,”
American Economic Journal: Macroeconomics, Forthcoming.
Chari, V. V., Kehoe, P. J., and McGrattan, E. R. (2009), “New Keynesian Models: Not Yet Useful
for policy Analysis,” American Economic Journal: Macroeconomics, 1(1), 242–266.
Christiano, L. J., Eichenbaum, M. S., and Trabandt, M. (2018), “On DSGE Models,” Journal of
Economic Perspectives, 32(3), 113–140.
Christoffel, K., Coenen, G., and Warne, A. (2008), “The New Area-Wide Model of the Euro Area: A
Micro-Founded Open-Economy Model for Forecasting and Policy Analysis,” ECB Working Paper
Series No. 944.
Christoffel, K., Coenen, G., and Warne, A. (2011), “Forecasting with DSGE Models,” in M. P.
Clements and D. F. Hendry (Editors), The Oxford Handbook of Economic Forecasting, 89–127,
Oxford University Press, New York.
Coenen, G., Karadi, P., Schmidt, S., and Warne, A. (2018), “The New Area-Wide Model II: An
Updated Version of the ECB’s Micro-Founded Model for Forecasting and Policy Analysis with a
Financial Sector,” ECB Working Paper Series No. 2200.
Cole, S. J. and Milani, F. (2019), “The Misspecification of Expectations in New Keynesian Models:
A DSGE-VAR Approach,” Macroeconomic Dynamics, 23(3), 974–1007.
RE AL RE RE AL
baseline baseline alt: gap alt: σc alt: ρ
subst. elasticity (σc ) 1 1.07 1 1.14 1.07
[calibrated] [1.04, 1.09] [calibrated] [1.05, 1.27] [1.04, 1.09]
belief persistence (ρ) – 0.17 – – 0.10
[0.07, 0.28] [0.01, 0.25]
Wage markup (AR) 0.73 0.85 0.88 0.77 0.84
[0.56, 0.86] [0.80, 0.88] [0.74, 0.96] [0.63, 0.88] [0.80, 0.87]
Wage indexation 0.25 0.19 0.25 0.24 0.19
[0.12, 0.40] [0.10, 0.29] [0.12, 0.40] [0.12, 0.39] [0.10, 0.30]
Wage stickiness 0.62 0.54 0.48 0.59 0.55
[0.54, 0.70] [0.48, 0.61] [0.40, 0.59] [0.52, 0.67] [0.49, 0.62]
Price markup (AR) 0.19 0 0.42 0.20 0
[0.05, 0.35] [calibrated] [0.18, 0.59] [0.07, 0.36] [calibrated]
Price indexation 0.22 0.23 0.15 0.21 0.23
[0.09, 0.35] [0.14, 0.33] [0.06, 0.32] [0.09, 0.34] [0.15, 0.33]
Price stickiness 0.80 0.80 0.72 0.79 0.80
[0.76, 0.84] [0.76, 0.84] [0.64, 0.80] [0.74, 0.84] [0.76, 0.84]
Notes: Posterior mean estimates are reported along with the 5th and 95th percentiles from the posterior
distribution in brackets. The models are estimated with the random-walk Metropolis algorithm using 750,000
posterior draws, where the first 250,000 are treated as burn-in sample. The data from 1980Q1–1984Q4 are used
as a training sample for the Kalman filter state variables. The log marginal likelihood is estimated with the
modified harmonic mean estimator. The gap alternative model under RE uses the output gap specification from
AL, while the σc alternative model under RE estimates the inverse elasticity of intertemporal substitution for
constant labor σc using the prior σc ∼ N (1, 0.252 ). The ρ alternative model under AL uses a standard uniform
prior distribution for the belief coefficients persistence parameter.
Notes: The cdf-values from the Diebold-Mariano (DM) test for the null hypothesis of equal mean squared
forecast errors (MSFE) are calculated as in Harvey, Leybourne, and Newbold (1997), equation (9). The cdf-
values shown above are taken from the Student’s t-distribution with Nh − 1 degrees of freedom, with Nh being
the number of h-quarter-ahead forecasts, Nh = 76 − h. A cdf-value close to zero suggests that the predictions of
the RE model are better in an MSFE sense than the those of the AL model, and a value close to one that the
reverse case is supported.
Notes: The full sample covers the vintages 2001Q1–2019Q4. The cdf-values for the Amisano-Giacomini weighted
LR test are taken from its asymptotic normal distribution using 1 lag for the HAC estimator and equal weights;
see Amisano and Giacomini (2007) for details. The cdf-values for the Diebold-Mariano test are taken from the
Student’s t-distribution with Nh − 1 degrees of freedom, with Nh being the number of h-quarter-ahead forecasts,
Nh = 76 − h. The DM test statistic is based on the difference of minus the log scores rather than the difference
in mean-squared forecast errors as in Harvey et al. (1997), equation (9). A cdf-value close to zero suggests that
the predictions of the RE model are better in a log score sense than those of the AL model, and a value close to
one that the reverse case is supported.
Notes: The cdf-values from the Diebold-Mariano (DM) test for the null hypothesis of equal CRPS are calculated
as in Harvey et al. (1997), equation (9), with minus the CRPS/ES values replacing the mean-squared forecast
errors. The cdf-values shown above are taken from the Student’s t-distribution with Nh − 1 degrees of freedom,
with Nh being the number of h-quarter-ahead forecasts, Nh = 76 − h. A cdf-value close to zero suggests that the
predictions of the RE model are better in a CRPS/ES sense than those of the AL model, and a value close to
one that the reverse case is supported.
Notes: The Amisano-Geweke tests are described in the Online Appendix of Amisano and Geweke (2017).
The statistics above are based on p = 2 lags and q = 2 moments such that the asymptotic distribution is
χ24 . The PIT (BOT) tests concern the marginal (joint) nowcasts and forecasts.
-2
-4
-6
-100
likelihood (AL)
posterior kernel (AL)
-150 Laplace (AL)
likelihood (RE)
posterior kernel (RE)
-200
Laplace (RE)
-250
-300
-350
-400
-450
2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020
-300
MHM (AL)
Laplace (AL)
-320 MHM (RE)
Laplace (RE)
-340
-360
-380
-400
-420
-440
2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020
Notes: The log marginal likelihood is estimated with (i) the modified harmonic mean (MHM) estimator using
10,000 equally spaced posterior draws from the 500,000 post burn-in draws of the random-walk Metropolis sampler
with a burn-in sample of 250,000 draws, and (ii) the Laplace approximation for the joint posterior mode. The
log likelihood and the log posterior kernel are also estimated at the joint posterior mode of the respective annual
vintage, each taken from Q1.
1 1
0 0
-1 -1
-2 -2
-3 -3
-4 -4
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
0.5 0.5
0 0
-0.5 -0.5
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
Notes: The actual values are plotted as solid black lines. Recursively estimated posterior predictive mean values
of real GDP growth and inflation are plotted as solid red and blue lines under RE and AL, respectively, while
the dashed black lines are the recursive posterior mean estimates of steady-state real GDP growth and inflation.
Adaptive Learning
AL Steady-State
Rational Expectations
0.1
0.05
-0.05
-0.1
-0.15
-0.2
1980
1990 40
2000 30
20
2010
10
Dates Horizon
2020 0
0.01
-0.01
-0.02
-0.03
-0.04
-0.05
-0.06
-0.07
1980
1990 40
2000 30
20
2010
10
Dates Horizon
2020 0
-3 -2
-4 -2.5 -1.5
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
-1 -1 -1
-1 -1 -1
eight-quarter-ahead
RE
-0.5
AL
Alt. PLM
-1
-1.5
2000 2005 2010 2015 2020
4
1
2
0.5
0
0
-2
-0.5 -4
2015 2016 2017 2015 2016 2017
-0.5
2015 2016 2017
Notes: The unconditional forecasts of the baseline RE model are plotted as a solid red line, while the 90 % (70
%) equal-tails credible interval of these forecasts are shown as a dark-grey (light-grey) area. The unconditional
forecasts of the baseline AL model are given by a dash-dotted blue line, while the borders of the 90 % (70 %)
equal-tails credible interval are plotted as solid (dashed) blue lines. Finally, the conditional forecasts of the
baseline RE model are given by the dash-dotted green line, while the borders of its 90 % (70 %) equal-tails
credible interval are plotted as solid (dashed) green lines. The conditional forecast of inflation are based on
the actual values of the short-term nominal interest rate in 2015Q1–2016Q4 using the Waggoner and Zha (1999)
conditioning approach. GDP deflator inflation is measured in quarterly terms and the short-term nominal interest
rate in annual terms.
seven-quarter-ahead eight-quarter-ahead
RE
AL
-0.3 -0.3 Alt. PLM
-0.4 -0.4
-0.5 -0.5
-0.6 -0.6
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
A. Nowcasts
Rational Expectations Adaptive Learning
0.5 0.5
0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
B. One-quarter-ahead forecasts
0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Notes: The horizontal axis shows the 10 bins while the vertical axis shows the occurence frequency for the
estimated π’s. If these variables are uniformly distributed for a model, then the occurence in large samples is
0.10 for all bins.
The forward looking variables in the model can be extracted from zt in equation (1) by constructing
a 0-1 selection matrix S of dimension p × f and having rank f ≤ p such that
ztf = S ′ zt .
Note that S is made up of f distinct columns of Ip . The remaining p − f columns are denoted by
S⊥ such that the non-forward looking variables are given by
ztnf = S⊥
′
zt .
We can now define the matrix S̃, which we will employed to transform (re-order) the structural
equations and the state variables, as follows:
S̃ = S S⊥ ,
where S̃ −1 = S̃ ′ .
The order of the equations in the DSGE model is arbitary and for the transformation approach
below it is important that the order at least temporarily follows the order in which the forward
looking variables appear in z. Moreover, it is required that
rank H1 ≤ f,
i.e., the rank of H1 provides a lower bound for the number of forward looking variables that are
supported by the model.
Concerning the reordering of equations, each forward looking variable in the expectation term
should appear in an equation having the same order number as the variable itself has among z. To
this end, the matrix C is defined as a p × p matrix of rank p containing only unique rows from Ip .
This means that C satisfies C ′ C = CC ′ = Ip .27
Premultiplying the system in (1) by S̃ ′ C, it can be rewritten as follows
H̃−1 z̃t−1 + H̃0 z̃t + H̃1 Et z̃t+1 = D̃ηt , (A.1)
where z̃t = S̃ ′ zt , H̃i = S̃ ′ CHi S̃ for i = −1, 0, 1, and D̃ = S̃ ′ CD. This means that the equations
for the forward looking variables are ordered in the top f equations (rows) and the bottom p − f
equations (rows) are those for the non-forward looking variables. Furthermore, the former variables
are given in the first f rows of z̃t and the latter variables in the bottom p − f rows.
27 Let ι denote an f dimensional vector with integers giving the position of each forward looking variable in z.
f
Similarly, let ιe be an f dimensional vector with positions of the rows in H1 that are non-zero. This vector need
not be unique as more rows than f may contain non-zero elements. For such situations, it is only required that the
sub-matrix formed from H1 using the rows ιe and the columns ιf has rank f . Provided that this condition is met, C
is constructed by first setting it to Ip . Next, the rows ιe in C are replaced with the rows ιf from Ip . Last, the rows
ιf of C are replaced with ιe of Ip .
In the case of H̃1 , the f ×f sub-matrix S ′ CH1 S has full rank f , while the sub-matrix S⊥
′ CH S = 0.
1 ⊥
These results follow directly from the assumptions that ztf is forward looking and that ztnf is non-
forward looking. For the sub-matrix in the bottom right corner of H̃1 to be zero, we find that either
′ CH = 0.
CH1 S⊥ = 0 or S⊥ 1
For the case when CH1 S⊥ = 0, we find that the final p − f columns of H̃1 are zero and that only
the expected next period values of the forward looking variables appear in the p equations. There
is therefore no need for any further transformation of the DSGE model as
f
H̃1 Et z̃t+1 = H̃1,f Et zt+1 .
Based on this we can replace the expectations with the adaptive learning mechanism for the forward
looking variables when we solve the model.
′ CH = 0 is somewhat more complicated as we need to transform the
The second case with S⊥ 1
system by substituting for the expectations of the non-forward looking variables in the top f equa-
tions. To accomplish this, we note that the bottom p − f equations in (A.1) do not involve any
expectations, but only contemporaneous and lagged variables. Under the assumption that the model
has a solution, it follows that
′
rank S⊥ CH0 S⊥ = p − f.
Accordingly, the solution of the DSGE model includes the following representation for the non-
forward looking variables
−1 −1 h i
ztnf = − S⊥
′
CH0 S⊥ ′
S⊥ CH0 Sztf − S⊥
′
CH0 S⊥ ′
S⊥ ′
CH−1 S S⊥ CH−1 S⊥ z̃t−1
′
−1 ′
+ S⊥ CH0 S⊥ S⊥ Dηt .
From this equation we see that the unbiased expectation of the non-forward looking variables at
t + 1 based on information at t is
nf ′
−1 ′ f ′
−1 h ′ ′
i
Et zt+1 = − S⊥ CH0 S⊥ S⊥ CH0 SEt zt+1 − S⊥ CH0 S⊥ S⊥ CH−1 S S⊥ CH−1 S⊥ z̃t .
H̄−1 z̃t−1 + H̄0 z̃t + H̄1 Et z̃t+1 = D̃ηt , (A.2)
At this stage, it is useful to premultiply the structural form by C ′ S̃ and replace z̃t and z̃t−1 with
zt and zt−1 , respectively, while the expectations term is kept. This provides us with
∗
f
H−1 zt−1 + H0∗ zt + H1,f
∗
Et zt+1 = Dηt . (A.3)
and
C ′ S̃ H̃1,f = H1 S
if CH1 S⊥ = 0,
∗
H1,f =
C ′ S̃ H̄
′ CH = 0.
if S⊥
1,f 1
These conditions and transformations are simple to apply once the forward looking variables have
been established. The case when all columns of H1 that are multiplied by a non-forward looking
variable are zero is very easy to spot and require hardly any rewrite of the DSGE model. The second
case when all rows of H1 in the equations for the non-forward looking variables are zero, require a
little bit more work, but is swiftly dealt with by computer code.
Table B.1: Linking the vintages of the RTD to the AWM updates.
RTD RTD common AWM AWM AWM RTD euro AWM euro
vintages start date update start date end date area concept area concept
2001Q1–2001Q4 1994Q1 2 1970Q1 1999Q4 12 12
2002Q1–2003Q2 1994Q1 3 1970Q1 2000Q4 12 12
2003Q3–2004Q2 1994Q1 4 1970Q1 2002Q4 12 12
2004Q3–2005Q3 1994Q1 5 1970Q1 2003Q4 12 12
2005Q4–2006Q2 1995Q1 5 1970Q1 2003Q4 12 12
2006Q3–2006Q4 1995Q1 6 1970Q1 2005Q4 12 12
2007Q1–2007Q2 1996Q1 6 1970Q1 2005Q4 12,13 12
2007Q3 1996Q1 7 1970Q1 2006Q4 12,13 13
2007Q4–2008Q1 1996Q1 7 1970Q1 2006Q4 13 13
2008Q2 1995Q1 7 1970Q1 2006Q4 13,15 13
2008Q3–2008Q4 1995Q1 8 1970Q1 2007Q4 15 15
2009Q1–2009Q2 1995Q1 8 1970Q1 2007Q4 16 15
2009Q3–2010Q2 1995Q1 9 1970Q1 2008Q4 16 16
2010Q3–2010Q4 1995Q1 10 1970Q1 2009Q4 16 16
2011Q1 1995Q1 10 1970Q1 2009Q4 16,17 16
2011Q2 1995Q1 10 1970Q1 2009Q4 17 16
2011Q3–2012Q2 1995Q1 11 1970Q1 2010Q4 17 17
2012Q3–2013Q2 1995Q1 12 1970Q1 2011Q4 17 17
2013Q3–2013Q4 1995Q1 13 1970Q1 2012Q4 17 17
2014Q1–2014Q2 2000Q1 13 1970Q1 2012Q4 18 17
2014Q3–2014Q4 2000Q1 14 1970Q1 2013Q4 18 18
2015Q1–2015Q2 2000Q1 14 1970Q1 2013Q4 19 18
2015Q3–2016Q1 2000Q1 15 1970Q1 2014Q4 19 19
2016Q2 1998Q1 15 1970Q1 2014Q4 19 19
2016Q3–2017Q2 1998Q1 16 1970Q1 2015Q4 19 19
2017Q3–2018Q2 1998Q2 17 1970Q1 2016Q4 19 19
2018Q3–2020Q4 1998Q2 18 1970Q1 2017Q4 19 19
Notes: Data from the AWM is always taken from 1980Q1 until the quarter prior to the RTD common start date.
When two RTD euro area concepts are indicated it means that some variables are based on one of them, while
others are based on the other. In all cases, the lower euro area country concept concerns unit labor cost, while the
higher concept is used for the aggregation of the other variables except in 2011Q1 when also the GDP deflator, total
employment and the unemployment rate is based on euro area 16. Unit labor cost is the measure underlying the
calculation of nominal wages as unit labor cost times real GDP divided by total employment. Unemployment has
undergone gradual changes in the definition in December 2000, March and June 2002; see, e.g., European Central Bank
(2001) for details.
Date y c i π e w r u
Backcast 2001Q1– 2001Q1– 2001Q1– 2001Q1– 2001Q1– 2001Q1– – –
2001Q2 2001Q2 2001Q2 2003Q3 2017Q3 2017Q7
2002Q1 2002Q1 2002Q1
2003Q3 2003Q3
2004Q3 2004Q3 2004Q3
2006Q1 2006Q1 2006Q1
2006Q3 2006Q3 2006Q3
2014Q3– 2014Q3– 2014Q3–
2015Q4 2015Q4 2015Q4
2016Q2 2016Q2 2016Q2
2018Q1 2018Q1
2019Q1 2019Q1 2019Q1 2019Q1
Total 3 of 76 15 of 76 15 of 76 22 of 76 68 of 76 69 of 76 0 of 76 0 of 76
Nowcast 2001Q1– 2001Q1– 2001Q1– 2001Q1– 2001Q1– 2001Q1– – 2005Q1
2019Q4 2019Q4 2019Q4 2019Q4 2019Q4 2019Q4 2005Q3–
2005Q4
2006Q3
2008Q1
Total 76 of 76 76 of 76 76 of 76 76 of 76 76 of 76 76 of 76 0 of 76 5 of 76
RE AL
parameter density P1 P2 P1 P2
ϕ N 4.0 1.5 4.0 1.5
σc N 1.0c – 1.0 0.25
λ β 0.7 0.1 0.7 0.1
σl N 2.0 0.75 2.0 0.75
ξw β 0.5 0.05 0.5 0.05
ξp β 0.5 0.1 0.5 0.1
ıw β 0.5 0.15 0.5 0.15
ıp β 0.5 0.15 0.5 0.15
φp N 1.25 0.125 1.25 0.125
ψ β 0.5 0.15 0.5 0.15
ρmp β 0.75 0.1 0.75 0.1
rπ N 1.5 0.25 1.5 0.25
ry N 0.125 0.05 0.125 0.05
r∆y N 0.125 0.05 0.125 0.05
ξe β 0.5 0.15 0.5 0.15
υ β 0.2 0.05 0.2 0.05
π̄ Γ 0.625 0.1 0.625 0.1
β̄ Γ 0.25 0.1 0.25 0.1
ē N 0.2 0.05 0.2 0.05
γ̄ N 0.3 0.05 0.3 0.05
α N 0.3 0.05 0.3 0.05
Notes: The columns P1 and P2 refer to the mean and the standard deviation of the normal (N ), standardized
beta (β), and gamma (Γ) distributions. The superscript c means that the parameter is calibrated. The parameter
ρmp is the coefficient on the lagged interest rate in the monetary plicy rule.
RE AL
parameter density P1 P2 P1 P2
ρ β – – 0.25 0.1
ρg β 0.5 0.2 0.5 0.2
ρga β 0.5 0.2 0.5 0.2
ρb β 0.5 0.2 0.5 0.2
ρi β 0.5 0.2 0.5 0.2
ρa β 0.5 0.2 0.5 0.2
ρp β 0.5 0.2 0.0c –
ρw β 0.5 0.2 0.5 0.2
ρr β 0.5 0.2 0.5 0.2
ρs β 0.5 0.2 0.5 0.2
σg U 0 5 0 5
σb U 0 5 0 5
σi U 0 5 0 5
σa U 0 5 0 5
σp U 0 5 0 5
σw U 0 5 0 5
σr U 0 5 0 5
σs U 0 5 0 5
Notes: The columns P1 and P2 refer to the mean and the standard deviation of the standardized beta distribution
and the upper and lower bound of the uniform (U ) distribution. The superscript c means that the parameter is
calibrated.
RE AL
mean mode 5% 95% mean mode 5% 95%
ϕ 5.03 4.86 3.85 6.41 7.50 7.40 6.03 9.03
σc 1.0c 1.0c – – 1.07 1.06 1.04 1.09
λ 0.64 0.65 0.57 0.70 0.87 0.87 0.84 0.89
σl 5.41 5.40 5.16 5.68 5.38 5.32 4.96 5.82
ξw 0.62 0.62 0.54 0.70 0.54 0.54 0.48 0.61
ξp 0.80 0.79 0.76 0.84 0.80 0.80 0.76 0.84
ıw 0.25 0.21 0.12 0.40 0.19 0.18 0.10 0.29
ıp 0.22 0.20 0.09 0.36 0.23 0.22 0.14 0.33
φp 1.54 1.55 1.41 1.67 1.36 1.35 1.23 1.50
ψ 0.52 0.52 0.40 0.64 0.58 0.58 0.37 0.77
ρmp 0.87 0.88 0.84 0.90 0.93 0.93 0.90 0.94
rπ 1.42 1.39 1.19 1.66 1.87 1.86 1.59 2.18
ry 0.19 0.19 0.15 0.24 0.08 0.08 0.04 0.14
r∆y 0.02 0.20 0.00 0.04 0.03 0.03 0.01 0.04
ξe 0.69 0.69 0.65 0.73 0.81 0.81 0.80 0.83
υ 0.16 0.16 0.08 0.25 0.04 0.03 0.02 0.06
π̄ 0.58 0.58 0.46 0.69 0.58 0.57 0.46 0.71
β̄ 0.22 0.21 0.11 0.35 0.24 0.23 0.12 0.40
ē 0.18 0.18 0.17 0.19 0.16 0.16 0.15 0.17
γ̄ 0.14 0.14 0.10 0.18 0.19 0.19 0.16 0.21
α 0.23 0.23 0.20 0.26 0.24 0.24 0.21 0.29
Notes: The columns for each model display the mean, the mode, and the 5% and 95% quantiles, respectively,
from the posterior distributions. The superscript c means that the parameter is calibrated.
RE AL
mean mode 5% 95% mean mode 5% 95%
ρ – – – – 0.17 0.15 0.07 0.28
ρg 0.99 0.99 0.99 0.99 0.98 0.99 0.98 0.99
ρga 0.25 0.25 0.16 0.36 0.25 0.25 0.16 0.35
ρb 0.92 0.92 0.89 0.95 0.36 0.37 0.23 0.48
ρi 0.15 0.14 0.06 0.27 0.14 0.12 0.04 0.25
ρa 0.98 0.98 0.98 0.99 0.93 0.94 0.91 0.95
ρp 0.19 0.15 0.05 0.35 0.0c 0.0c – –
ρw 0.73 0.77 0.56 0.86 0.85 0.85 0.80 0.89
ρr 0.29 0.29 0.18 0.41 0.33 0.32 0.22 0.44
ρs 0.98 0.98 0.97 0.99 0.98 0.98 0.97 0.99
σg 0.31 0.30 0.28 0.34 0.30 0.30 0.27 0.34
σb 0.05 0.05 0.04 0.06 0.11 0.11 0.08 0.14
σi 0.55 0.55 0.48 0.63 0.26 0.25 0.20 0.33
σa 0.49 0.48 0.41 0.60 0.54 0.52 0.44 0.66
σp 0.24 0.21 0.14 0.38 0.04 0.03 0.03 0.05
σw 0.12 0.09 0.06 0.23 0.06 0.06 0.05 0.07
σr 0.10 0.10 0.09 0.11 0.09 0.09 0.08 0.10
σs 1.02 1.01 0.92 1.14 0.82 0.81 0.72 0.93
∆y π ∆y, π
h RE AL RE AL RE AL
0 0.40 0.32 0.40 0.16 0.40 0.24
1.00 1.00 1.00 1.00 1.00 1.00
1 0.49 0.33 0.33 0.50 0.49 0.41
0.97 1.00 1.00 0.97 0.97 1.00
2 0.33 0.33 0.74 0.25 0.66 0.33
1.00 1.00 0.64 1.00 0.78 1.00
3 0.41 0.41 0.50 0.41 0.41 0.25
1.00 1.00 0.97 1.00 1.00 1.00
4 0.42 0.42 0.42 0.41 0.50 0.42
1.00 1.00 1.00 1.00 0.96 1.00
5 0.67 0.42 0.59 0.42 0.50 0.34
0.76 0.99 0.88 0.99 0.96 1.00
6 0.59 0.42 0.34 0.42 0.42 0.42
0.88 0.99 1.00 0.99 0.99 0.99
7 0.68 0.51 0.34 0.43 0.42 0.51
0.74 0.96 1.00 0.99 0.99 0.96
8 0.69 0.34 0.34 0.77 0.34 0.60
0.73 1.00 1.00 0.59 1.00 0.86
Notes: Real GDP growth is denoted by ∆y and GDP deflator inflation by π. The Kolmogorov-
Smirnov test is here computed as Nh /2 · supxi |FN1 h (xi ) − FN2 h (xi )|, where FNj h is the empirical
p
cumulative distribution function of the log predictive likelihood using estimator j, while Nh is
the number of predictive likelihood values of the h-quarter-ahead forecasts. The mean and the
standard deviation of the Kolmogorov distribution are roughly 0.87 and 0.26, respectively. The
critical values of the test statistic for sizes 10, 5 and 1 percent are about 1.22, 1.36 and 1.63,
p
which may be calculated from cα = −(1/2) log(α/2)), with α being the size of the test and cα
the critical value. The p-value of the test statistic is shown below the test value and has been
computed using a truncation of 1,000 for an expression of its limiting distribution; see, for instance,
Marsaglia, Tsang, and Wang (2003).
Figure C.1: The observed variables from the full sample, 1985Q1–2019Q4.
2
1 1
0
0 0
-2
-1 -1
-4
-2 -2 -6
1990 2000 2010 2020 1990 2000 2010 2020 1990 2000 2010 2020
1
0.5
1
0.5
0
0
0
-0.5
-0.5
-1 -1 -1
1990 2000 2010 2020 1990 2000 2010 2020 1990 2000 2010 2020
10 10
5 5
0
0
1990 2000 2010 2020 1990 2000 2010 2020
0 0 0
-1 -1 -1
-2 -2 -2
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
0 0 0
-1 -1 -1
-2 -2 -2
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
0 0 0
-1 -1 -1
-2 -2 -2
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
eight-quarter-ahead
1 RE
AL
-1
-2
2000 2005 2010 2015 2020
0 0 0
-1 -1 -1
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
0 0 0
-1 -1 -1
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
0 0 0
-1 -1 -1
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
eight-quarter-ahead
1 RE
AL
0.5
-0.5
-1
2000 2005 2010 2015 2020
-0.5
-0.6
-1 -1
-0.7
-0.8
-1 -1 -1
-1 -1 -1
eight-quarter-ahead
-0.5 RE
AL
-1
-1.5
2000 2005 2010 2015 2020
0
0.2
0
-0.5
0
-1
-1
-0.2
-1.5
-2 -2 -0.4
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
0 0 0
0 0 0
eight-quarter-ahead
0.4 RE
AL
0.2
-0.2
-0.4
2000 2005 2010 2015 2020
seven-quarter-ahead eight-quarter-ahead
-0.2 -0.2
RE
AL
-0.25 -0.25
-0.3 -0.3
-0.35 -0.35
-0.4 -0.4
-0.45 -0.45
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
seven-quarter-ahead eight-quarter-ahead
RE
AL
-0.1 -0.1
-0.15 -0.15
-0.2 -0.2
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020
A. Nowcasts
Rational Expectations Adaptive Learning
0.5 0.5
0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
B. One-quarter-ahead forecasts
0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Notes: The horizontal axis shows the 10 bins while the vertical axis shows the occurence frequency for the
estimated π’s. If these variables are uniformly distributed for a model, then the occurence in large samples is
0.10 for all bins.
A. Nowcasts
Rational Expectations Adaptive Learning
0.5 0.5
0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
B. One-quarter-ahead forecasts
0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Anders Warne
European Central Bank, Frankfurt am Main, Germany; email: anders.warne@ecb.europa.eu